[Intel-gfx] [PATCH 2/2] drm/i915: add syncobj timeline support

[Lionel Landwerlin]

On 23/05/2019 14:59, Chris Wilson wrote:
> Quoting Lionel Landwerlin (2019-05-23 14:46:42)
>> On 23/05/2019 12:52, Chris Wilson wrote:
>>> Quoting Lionel Landwerlin (2019-05-23 12:46:20)
>>>> -               syncobj = drm_syncobj_find(file, fence.handle);
>>>> -               if (!syncobj) {
>>>> -                       DRM_DEBUG("Invalid syncobj handle provided\n");
>>>> -                       err = -ENOENT;
>>>> -                       goto err;
>>>> +                       if (user_fence.flags & __I915_EXEC_FENCE_UNKNOWN_FLAGS) {
>>>> +                               err = -EINVAL;
>>>> +                               goto err;
>>>> +                       }
>>>> +
>>>> +                       if (user_fence.flags & I915_EXEC_FENCE_WAIT) {
>>>> +                               err = drm_syncobj_find_fence(
>>>> +                                       file, user_fence.handle, user_fence.value,
>>>> +                                       DRM_SYNCOBJ_WAIT_FLAGS_WAIT_FOR_SUBMIT,
>>>> +                                       &syncobj, &fence);
>>> Is this still a synchronous wait? That would be an unfortunate change in
>>> behaviour and antithesis to having a scheduler.
>>> -Chris
>>>
>> Not sure what you mean by synchronous wait.
> drm_syncobj_find_fence() has an open-coded wait_event loop. That is
> synchronous and inconsistent with using a scheduler; where one only need
> to return a proxy fence that will be populated when the syncpt is known,
> and be signaled as a result of that syncpt.
> -Chris
>

Right,


I see this changes the behavior for existing drm_syncobjs, which was not 
intended.


For timeline drm-syncobjs, we're aware this will block and might fail if 
the fence doesn't materialize.

The Vulkan spec doesn't give a specific amount of time of which the 
submit fails, apart from it being non zero.


Will fix, thanks a lot for pointing this out.


-Lionel